r70

This release includes bugfix for faster rendering with slow gpus. Also scrolling heatmap view of neural network weights was added to many demos.

r53

All ConvNetJS, ReinforceJS and RecurrentJS examples are translated to this release. Every example has working continuous functionality and loading of the pretrained settings.

User is able to get programs to crash, if user creates invalid JSON files. Error handling is not perfect, but I consider this to be comparable to a v1.0.0 release, as correct usage works as in original examples.

I guess that the highlight for most people in this release is the CIFAR-10 classification example. In the other hand, I am personally very happy about the implementation of the LSTM network, which avoided all heap memory allocations during busy loop with pre-created structure, while the original relied heavily on dynamic recursive structures.

I have tried to optimize the performance. Busy loops doesn't generally do heap allocations after first allocation.